Ensure that the driver can listen to evictions even when they don't take the
path through ttm_bo_driver::move.

This is crucial for amdgpu, which relies on an eviction counter to skip
re-binding page tables when possible.
Signed-off-by: NNicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: NChunming Zhou <david1.zhou@amd.com>
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

TTM was changed a while back to allow for pipelining of buffer moves, and
part of this was the removal of waiting for a BO to idle before calling
move(), placing the responsibility on the driver to do this if required.

That's all well and good, except, we make use of move_notify() to handle
mapping/unmapping from the GPU VMM as move() isn't called on all paths.

This commit adds a wait before unmapping from a VMM in move_notify(), to
prevent GPU page faults where a buffer is still being accessed.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>
Cc: stable@vger.kernel.org [v4.8+]

This way the driver can decide if it is valuable to evict a BO or not.

The current implementation is added as default to all existing drivers.

v2: fix some typos found during internal testing
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

I plan to usurp the short name of struct fence for a core kernel struct,
and so I need to rename the specialised fence/timeline for DMA
operations to make room.

A consensus was reached in
https://lists.freedesktop.org/archives/dri-devel/2016-July/113083.html
that making clear this fence applies to DMA operations was a good thing.
Since then the patch has grown a bit as usage increases, so hopefully it
remains a good thing!

(v2...: rebase, rerun spatch)
v3: Compile on msm, spotted a manual fixup that I broke.
v4: Try again for msm, sorry Daniel

coccinelle script:
@@

@@
- struct fence
+ struct dma_fence
@@

@@
- struct fence_ops
+ struct dma_fence_ops
@@

@@
- struct fence_cb
+ struct dma_fence_cb
@@

@@
- struct fence_array
+ struct dma_fence_array
@@

@@
- enum fence_flag_bits
+ enum dma_fence_flag_bits
@@

@@
(
- fence_init
+ dma_fence_init
|
- fence_release
+ dma_fence_release
|
- fence_free
+ dma_fence_free
|
- fence_get
+ dma_fence_get
|
- fence_get_rcu
+ dma_fence_get_rcu
|
- fence_put
+ dma_fence_put
|
- fence_signal
+ dma_fence_signal
|
- fence_signal_locked
+ dma_fence_signal_locked
|
- fence_default_wait
+ dma_fence_default_wait
|
- fence_add_callback
+ dma_fence_add_callback
|
- fence_remove_callback
+ dma_fence_remove_callback
|
- fence_enable_sw_signaling
+ dma_fence_enable_sw_signaling
|
- fence_is_signaled_locked
+ dma_fence_is_signaled_locked
|
- fence_is_signaled
+ dma_fence_is_signaled
|
- fence_is_later
+ dma_fence_is_later
|
- fence_later
+ dma_fence_later
|
- fence_wait_timeout
+ dma_fence_wait_timeout
|
- fence_wait_any_timeout
+ dma_fence_wait_any_timeout
|
- fence_wait
+ dma_fence_wait
|
- fence_context_alloc
+ dma_fence_context_alloc
|
- fence_array_create
+ dma_fence_array_create
|
- to_fence_array
+ to_dma_fence_array
|
- fence_is_array
+ dma_fence_is_array
|
- trace_fence_emit
+ trace_dma_fence_emit
|
- FENCE_TRACE
+ DMA_FENCE_TRACE
|
- FENCE_WARN
+ DMA_FENCE_WARN
|
- FENCE_ERR
+ DMA_FENCE_ERR
)
 (
 ...
 )
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NGustavo Padovan <gustavo.padovan@collabora.co.uk>
Acked-by: NSumit Semwal <sumit.semwal@linaro.org>
Acked-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/20161025120045.28839-1-chris@chris-wilson.co.uk

This reverts commit aff51175.

The commit caused fence timeouts within nvc0_screen_destroy and most likely
other places as well.

The most obvious effect is, that userspace processes take minutes to
actually quit.
Signed-off-by: NKarol Herbst <karolherbst@gmail.com>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Rather than using "struct file*", use "struct drm_file*" as tag VM tag for
BOs. This will pave the way for "struct drm_file*" without any "struct
file*" back-pointer.
Signed-off-by: NDavid Herrmann <dh.herrmann@gmail.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/20160901124837.680-3-dh.herrmann@gmail.com

Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NMichel Dänzer <michel.daenzer@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NMichel Dänzer <michel.daenzer@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Fixes hangs under memory pressure, e.g. running the piglit test
tex3d-maxsize concurrently with other tests.

Fixes: 17d33bc9 ("drm/ttm: drop waiting for idle in ttm_bo_evict.")
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NMichel Dänzer <michel.daenzer@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Fixes hangs under memory pressure, e.g. running the piglit test
tex3d-maxsize concurrently with other tests.

Fixes: 17d33bc9 ("drm/ttm: drop waiting for idle in ttm_bo_evict.")
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NMichel Dänzer <michel.daenzer@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

This flag's only remaining function is to ignore the uncached flag for
BOs on coherent architectures.

However the reason for allocating an object uncache on a non-coherent
architecture (namely because the cost of doing explicit flushes/
invalidations is higher than the benefit of caching the data because
accesses are few and far between) should also apply on architectures for
which coherency is maintained implicitly. Thus allocate coherent objects
as uncached on all architectures.
Signed-off-by: NAlexandre Courbot <acourbot@nvidia.com>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

TTM-allocated coherent objects were populated using the DMA API and
accessed using the mapping it returned to workaround coherency
issues. These issues seem to have been solved, thus remove this extra
case to handle and use the regular kernel mapping functions.
Signed-off-by: NAlexandre Courbot <acourbot@nvidia.com>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

It isn't used and not waiting for the GPU after scheduling a move is
actually quite dangerous.
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

When we want to pipeline accelerated moves we need to wait in the fallback path.
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Wait for idle before moving the BO in all drivers implementing
an accelerated move function.

This should keep the current behavior when removing the pre move wait.
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Fixes out-of-tree build issue where uapi/drm/nouveau_drm.h gets picked
up instead.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

This allows fine grained control for the driver where to add a BO into the LRU.

v2: fix typo in comment
Reviewed-by: NSinclair Yeh <syeh@vmware.com>
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Not used any more.
Reviewed-by: NSinclair Yeh <syeh@vmware.com>
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

Not used any more.
Reviewed-by: NSinclair Yeh <syeh@vmware.com>
Signed-off-by: NChristian König <christian.koenig@amd.com>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

The DMA API has different semantics on different architectures.
Currently on arm64, it can only provide memory from a small pool which
dries up quickly if we attempt to allocate big buffers from it.

Do not consider that option when running on non-x86, since regular TTM
buffers are the (current) best-fit for ARM platforms.
Signed-off-by: NAlexandre Courbot <acourbot@nvidia.com>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

We need to use post-decrement to get the dma_map_page undone also for
i==0, and to avoid some very unpleasant behaviour if dma_map_page
failed already at i==0.
Signed-off-by: NRasmus Villemoes <linux@rasmusvillemoes.dk>
Reviewed-by: NBen Skeggs <bskeggs@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

We already express the drm/agp depencies correctly in Kconfig, so we
can rip this remnant from the shared drm core days.

Aside: Pretty much all the #ifdefs in radeon/nouveau could be killed
if ttm would provide dummy functions. I'm not going to volunteer for
that though.

v2: Use IS_ENABLED(CONFIG_AGP) as suggested by Ville

v3: Polish from Ville's review.

Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Christian König <christian.koenig@amd.com> (v2)
Reviewed-by: NDavid Herrmann <dh.herrmann@gmail.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

This commit reinstates the pre-DEVINIT AGP fiddling that was broken in
an earlier commit.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

A variety of tweaks to the NVIF library interfaces, mostly ripping out
things that turned out to be not so useful.

- Removed refcounting from nvif_object, callers are expected to not be
  stupid instead.
- nvif_client is directly reachable from anything derived from nvif_object,
  removing the need for heuristics to locate it
- _new() versions of interfaces, that allocate memory for the object
  they construct, have been removed.  The vast majority of callers used
  the embedded _init() interfaces.
- No longer storing constructor arguments (and the data returned from
  nvkm) inside nvif_object, it's more or less unused and just wastes
  memory.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

This is purely preparation for upcoming commits, there should be no
code changes here.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

The namespace of NVKM is being changed to nvkm_ instead of nouveau_,
which will be used for the DRM part of the driver.  This is being
done in order to make it very clear as to what part of the driver a
given symbol belongs to, and as a minor step towards splitting the
DRM driver out to be able to stand on its own (for virt).

Because there's already a large amount of churn here anyway, this is
as good a time as any to also switch to NVIDIA's device and chipset
naming to ease collaboration with them.

A comparison of objdump disassemblies proves no code changes.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

NVKM is having it's namespace switched to nvkm_, which will conflict
with these functions (which are workarounds for the fact that as of
yet, we still aren't able to split DRM and NVKM completely).

A comparison of objdump disassemblies proves no code changes.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Switch to NVIDIA's name for the device.

The namespace of NVKM is being changed to nvkm_ instead of nouveau_,
which will be used for the DRM part of the driver.  This is being
done in order to make it very clear as to what part of the driver a
given symbol belongs to, and as a minor step towards splitting the
DRM driver out to be able to stand on its own (for virt).

Because there's already a large amount of churn here anyway, this is
as good a time as any to also switch to NVIDIA's device and chipset
naming to ease collaboration with them.

A comparison of objdump disassemblies proves no code changes.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Remove the function nouveau_bo_rd16() that is not used anywhere.

This was partially found by using a static code analysis program
called cppcheck.
Signed-off-by: NRickard Strandqvist <rickard_strandqvist@spectrumdigital.se>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

nouveau_ttm_tt_unpopulate() is supposed to return right after calling
ttm_dma_unpopulate() in the case of a coherent buffer. The return
statement was omitted, leading to the pages being unmapped twice. Fix
this.
Signed-off-by: NAlexandre Courbot <acourbot@nvidia.com>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>